numpy手写CNN卷积

第一步：先写一个只有待卷积矩阵data和卷积核k的方法，默认步长step==1,不设置padding

import numpy as np
def convolution(data, k):
    x, y = k.shape
    n, m = data.shape
    result = []
    for i in range(n - x + 1):
        line = []
        for j in range(m - y + 1):
            a = data[i:i + x, j:j + y]
            line.append(np.sum(np.multiply(k, a)))
        result.append(line)
    return np.array(result)

第二步：补充padding功能，默认设置上下左右一层

def convolution(data, k, padding):
    x, y = k.shape
    n, m = data.shape
    result = []
    if padding:
        data_0 = np.pad(data, ((1, 1), (1, 1)), 'constant', constant_values=(0, 0))
        print(data_0)
        for i in range(n):
            line = []
            for j in range(m):
                a = data_0[i:i + x, j:j + y]
                line.append(np.sum(np.multiply(k, a)))
            result.append(line)
    else:
        for i in range(n - x + 1):
            line = []
            for j in range(m - y + 1):
                a = data[i:i + x, j:j + y]
                line.append(np.sum(np.multiply(k, a)))
            result.append(line)
    return np.array(result)

第三步：补充步长step功能

def convolution(data, k, step, padding):
    x, y = k.shape
    n, m = data.shape
    result = []
    if padding:
        data_0 = np.pad(data, ((1, 1), (1, 1)), 'constant', constant_values=(0, 0))
        print(data_0)
        for i in range(n//step):
            line = []
            for j in range(m//step):
                a = data_0[i:i + x, j:j + y]
                line.append(np.sum(np.multiply(k, a)))
            result.append(line)
    else:
        for i in range((n - x)//step + 1):
            line = []
            for j in range((m - y)//step + 1):
                a = data[i:i + x, j:j + y]
                line.append(np.sum(np.multiply(k, a)))
            result.append(line)
    return np.array(result)

第四步：将padding设置为自适应，不用硬编码

def convolution(data, k, step, padding):

    x, y = k.shape
    n, m = data.shape
    result = []
    if padding:
        padding_h = (x - 1) // 2
    	padding_v = (y - 1) // 2
        data_0 = np.pad(data, ((padding_h, padding_h), (padding_v, padding_v)), 'constant',
                        constant_values=(0, 0))
        print(data_0)
        for i in range((n - x + 2 * padding_h) // step + 1):
            line = []
            for j in range((m - y + 2 * padding_v) // step + 1):
                a = data_0[i:i + x, j:j + y]
                line.append(np.sum(np.multiply(k, a)))
            result.append(line)
    else:
        for i in range((n - x) // step + 1):
            line = []
            for j in range((m - y) // step + 1):
                a = data[i:i + x, j:j + y]
                line.append(np.sum(np.multiply(k, a)))
            result.append(line)
    return np.array(result)

这写主要针对二维图像的卷积，后续有空继续更新。